Benchmarking Explainable AI Methods for Tabular Healthcare Data:Towards Standardized and Clinically Interpretable Evaluation Frameworks

Abstract

Artificial Intelligence (AI) is becoming more common in healthcare, but clinical confidence in AI is limited by black-box models. This research study evaluates Explainable AI (XAI) on the Pima Indians Diabetes database using Logistic Regression and Random Forest,coupled with SHAP,LIME,and Permutation Feature Importance (PFI) methods. Overall,Random Forest performance yielded an accuracy score of 84 percent and an ROC-AUC score of 0.80, while glucose,BMI,and age were consistently important predictors. SHAP identifies patterns. LIME explored the patterns for individuals and PFI confirmed importance scoring. In comparison to previous studies that have evaluated explainers in isolation,this study offers a framework that combined methods to improve on both accuracy and interpretability,and establish more meaningful, transparent and clinically-useful AI systems.

Ähnliche Arbeiten